Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Defense

Army picks AWS and WWT for Project Griffin autonomous cyber defense pilot

The U.S. Army has selected Amazon Web Services and World Wide Technology for Project Griffin, a pilot worth a combined $2.6 million to build an autonomous, agentic cyber defense capability for Army networks. The selection was made on 23 October and announced by Army Public Affairs on 26 October, according to the Army.

Griffin is one of three accelerated efforts under Project ARDS (Army Rapid Defensive of Cyber Systems), led by the Army’s principal cyber advisor. It arrives as Army Cyber Command reports that it has already built a full team of AI agents for cyber protection work and handed it to the joint force.

/ 01What happened

The Army describes Griffin as an effort to give cyber defenders an agentic platform that neutralises threats faster, and to reduce vulnerabilities caused by manual data analysis and human action. Project Manager, Cyber Warfare, will run execution. Brandon Pugh, principal cyber advisor to the Secretary of the Army, said the project “moves the Army’s superintelligence-cyber efforts into concrete action,” and that the Army will “swiftly implement SI-enabled cyber tools to detect, divert and deceive threat actors.”

Pugh had said earlier that the service received more than 100 industry submissions and narrowed them to two, DefenseScoop reported on 12 October. He told Federal News Network that the selected companies will work on the response component, and that after Griffin the Army plans to move to deception and then to agentic red or purple teaming. Of autonomous response, he said: “That is extraordinary. Something we have not done and we are very excited about.”

In parallel, Army Cyber Command’s Task Force Lexington, stood up in April, has mapped out a complete cyber protection team of agents, with roles including analytic support, data engineering, network analysis and host analysis. Lt. Gen. Christopher Eubank, the ARCYBER commander, said the agents were trained and run in a sandbox before being allowed onto the Army’s networks, and that humans check their work and retrain them before moving to the next task. “We’re treating each one of these agents just like a human operator, so they are basic,” he said. He added that no agents have strayed so far, which he attributed to strict guardrails that will be loosened as agents mature. The team has been passed to the Department of Defense Cyber Defense Command.

/ 02Why it matters

Cyber defense is one of the first places where the military is moving from AI that advises to AI that acts. The reason is speed: attacks unfold faster than human teams can triage alerts. The Army’s own framing of Griffin treats manual analysis and human action as a source of vulnerability.

That makes the oversight model the most important detail. ARCYBER’s approach, treating agents like junior operators, starting them in a sandbox, checking their work and widening their permissions gradually, is a sensible pattern borrowed from how human cyber operators are trained. The Army’s Griffin announcement itself says little about human approval in the response loop. Federal News Network’s reporting describes the goal as autonomous response, which is precisely where the question of how actions are reviewed becomes pressing.

/ 03Autonomy needs a record of reasons

When a human defender isolates a host or blocks a range of addresses, they can explain why: this alert, correlated with that log entry, matched this known technique. When an agent does it, the same explanation has to exist, in a form someone can inspect after the fact. Otherwise a false positive that takes down a mission system is very hard to diagnose, and a successful response is hard to learn from.

A structured model of the network and the threat helps with both. If assets, users, services, vulnerabilities, indicators and adversary techniques are typed entities in a graph, an agent’s decision can be recorded as a path through that graph: the evidence it used, the rule or reasoning step it applied, and the action it took. That record is deterministic and replayable. It is also what lets operators widen an agent’s permissions with confidence, because they can see how it reasoned before it was trusted with more.

/ 04What to watch

  • The pilot’s rules of engagement. Which response actions Griffin agents may take without a human, and how those actions are logged.
  • Deception and red teaming. Pugh said funding is aligned for the follow-on phases.
  • Joint adoption. How the Department of Defense Cyber Defense Command uses ARCYBER’s agent team across other networks.
  • Failure reporting. Whether the Army publishes lessons from agent errors, which would help the wider community calibrate trust.
How it unfolded
APR 2026Task Force Lexingtonstood up[3]OCT 12, 2026Submissions narrowedto two[2]OCT 23, 2026AWS and WWT selected[1]OCT 26, 2026Army announcesGriffin partners[1]OCT 27, 2026This postAPR 2026Task Force Lexington stood up [3]OCT 12, 2026Submissions narrowed to two [2]OCT 23, 2026AWS and WWT selected [1]OCT 26, 2026Army announces Griffin partners [1]OCT 27, 2026This post
Aktiver’s take — Defense & intelligence

Autonomous response is a sensible place to start agentic cyber defense, and the wrong place to skip the paper trail. Every permission a Griffin agent gains should be earned with a record of how it used the last one.

Model the defended network as a graph: hosts, accounts, services, vulnerabilities, indicators and adversary techniques as typed entities. An agent that isolates a host then writes its decision as a path: alert A on host H, correlated with log entry L, matched technique T, therefore action X under rule R. Because reasoning over that graph is deterministic, a watch officer can replay the same inputs and get the same action, which turns a suspected false positive into a diagnosable one. It also gives the people loosening an agent’s guardrails something concrete to review: not a success rate, but every path it took while still in the sandbox. Knowledge Builder agents run with semantic enforcement over the graph, so an action with no supporting path fails loudly instead of quietly.

How we’d model it

Response actionAlert correlationAlertHostIndicatorAdversary techniqueAgentRules of engagementjustifiedByobservedOnmatchespermittedByResponseactionAlertcorrelationAlertHostIndicatorAdversarytechniqueAgentRules ofengagementjustifiedByobservedOnmatchespermittedBy
  • Response action —justifiedBy→ Alert correlation
  • Alert —observedOn→ Host
  • Indicator —matches→ Adversary technique
  • Agent —permittedBy→ Rules of engagement

See where this applies →

Sources — read the originals
  1. Army selects Project Griffin partners to accelerate SI-enabled cyber defense — U.S. Army army.mil ↗
  2. ARCYBER expanding use of AI agents for network operations — DefenseScoop defensescoop.com ↗
  3. Army Cyber Command builds agentic workforce to support cyber operations — Federal News Network federalnewsnetwork.com ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading